Research on Human Action Recognition Based on Global-Local Features of Video

Abstract

In the research of human behavior recognition, the two-stream network structure shows excellent results. Aiming at the branch feature of two-stream networks, this paper proposes a two-stream human behavior research method based on global-local features. This method first uses a mixture of Gaussian background modeling methods to extract silhouette features as global contour features, and then uses an end-to-end learnable unsupervised network TV-Net to generate optical flow motion features, which are used as the network input, and the Xception network is used as The feature generation network which does not change the model scale while improving the accuracy, and performs fusion classification on the output of the two-stream branch network to obtain the behavior recognition result. This method refines the motion information contained in the global and local features for classification, reduces the computational complexity, and shows a good level of recognition on both public and internal data sets.